Research On Medical Consumable Consumption Prediction Algorithm Based On NeuralProphet-GRU Model

Objective: Medical inventory management is an integral part of hospital management,especially with the increasing demand for precision medical care from governments and hospitals,the importance of medical inventory management has become more prominent.In this context,medical consumables usage prediction is even more important.However,traditional methods for predicting medical consumables usage,such as ARIMA and Neural Prophet,have not been able to effectively capture long sequence features.The purpose of this paper is to use Neural Prophet as a mixed method,combined with a GRU network that can capture long sequence features,to construct a new Neural Prophet-GRU model,so that it can more effectively capture the data characteristics of medical consumables inventory usage and improve the predictive performance.Methods: This paper first models the medical consumables prediction problem as a time series forecasting problem.To address this issue,the paper innovatively combines the GRU network with the Neural Prophet network to construct a new combination model,Neural Prophet-GRU.The model consists of neural network-based autoregressive terms,lag terms,and statistical trend,periodicity,event,and exogenous terms.The paper proposes an AR-GRU neural network that inputs historical values into the GRU hidden unit based on the input window length,and adaptively updates the hidden layer parameter values through a combination of two gating circuits: the update gate and the correlation gate.The AR-GRU is applied in both the autoregressive term module and the lag term module.The autoregressive term module models the future value of the target variable by autoregressively modeling its own features.The lag term uses the features of other related covariates to predict the future value of the target variable.The trend term models the overall trend of medical consumables using a continuous piecewise linear sequence.The periodicity term automatically identifies daily,weekly,and annual periodic features of the consumable sequence and models them using Fourier terms.The event term adjusts the weights for special events such as holidays to adjust deviations.The exogenous term uses known future external variables to assist in modeling the target variable.To use the optimal hyperparameters for training,this method also includes a training parameter self-adjustment function,which automatically selects the optimal parameters for all hyperparameters,including loss function,regularization,optimizer,learning rate,batch size,iteration number,and scheduler.The consumable data used in this paper are all from the billing details of medical consumables and diagnosis and treatment projects at Jin Qiu Hospital in Liaoning Province.Results: The validity of the model was verified through experiments that fully utilized the functions provided by the various modules of the algorithm,demonstrating that the algorithm has good interpretability and ease of use.The algorithm was tested on 7different types of medical consumables,achieving optimal predictive performance with a w MAPE of 0.162 for drug delivery syringes with a 90-day window length,and excellent results for large-scale samples.Comparative experiments were conducted between this algorithm and four other mainstream algorithms,showing that this algorithm can more effectively capture deep sequence information and achieve superior predictive performance.It outperformed all other algorithms when the prediction window length exceeded 30,achieving optimal predictive performance of 0.150 when the prediction window length was 90.This paper also conducted association prediction experiments,using the correlation between various consumables,drugs,and diagnosis and treatment projects to construct covariates to improve the processing ability of the target variables.The experiments showed that the model has strong representation ability for multiple types of correlated covariates,which can effectively improve the accuracy of prediction.The experiments also verified that the model’s ability to handle long sequences can be effectively applied to the annual budget and index declaration of hospitals,further improving the utilization of medical resources.Conclusion: This paper proposes the Neural Prophet-GRU model for medical consumables inventory prediction,achieving accurate forecasting of medical consumables consumption.The application of this algorithm in hospital consumables management systems can assist management personnel in predicting inventory consumption with precision,which has significant real-world implications for cost reduction and management optimization in hospitals.